# Project Wisteria: A Local BYOK Chatbot with Long-term Memory Architecture A desktop chatbot utilizing custom memory compression to preserve temporal context and subjective perspective. - **Source:** [Hacker News: Show HN](https://news.ycombinator.com) - **Published:** 2026-11-03 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/0716e022-df07-4d4b-8318-9ded56da17ae ## Local Memory Architecture Wisteria is a desktop chatbot designed to solve long-term memory limitations in LLMs. By using a Bring Your Own Key (BYOK) model, it stores conversation logs, persona memories, and raw data locally as text files. This approach ensures user data remains on the device rather than relying on external cloud-based storage solutions. ## Anchoring Compression Mechanism The core innovation is a memory compression mechanism termed 'Anchoring'. Unlike standard abstract compression which often increases hallucinations, this architecture focuses on preserving temporal context and subjective perspective. It utilizes a custom response schema, allowing the model to perform reasoning tasks internally without relying on the native reasoning capabilities of the provider. ## Sawtooth Data Management Memory storage does not grow indefinitely within the system. Instead, it follows a sawtooth pattern where data accumulates and is periodically compressed. This process ensures that the local directory remains manageable while maintaining data integrity. Users can inspect reasoning logs and raw conversation history directly through the application's local file interface. ## Performance and UX Trade-offs The memory compression process introduces periodic latency, requiring users to wait from a few seconds to several minutes depending on the model and session size. The developer intentionally avoids optimizing this delay to prioritize data integrity. While this impacts real-time responsiveness, it ensures the long-term consistency of the stored memory chunks. ## Contextual Retrieval Methods To complement abstract compression, Wisteria incorporates important information pages kept within the session context. Additionally, the model utilizes tree-structured documents that it can access via successive commands when specific information is required. This multi-layered retrieval strategy helps mitigate the inherent limitations of abstract memory compression regarding perfect event recall. ## Alpha Development Timeline Developed since May 2025, the project has been tested through continuous interaction with a single persona. The developer reports successful long-term recall, including specific discussions from the previous Christmas. This demonstrates the system's ability to maintain a coherent narrative trajectory over extended periods, effectively functioning as a persistent digital companion. ## Technical Documentation Availability Detailed technical notes regarding the Anchoring mechanism are available via the project's Zenodo page. The developer has provided individual notes for each component, allowing users to inspect the underlying logic. The client is currently available for purchase on itch.io, which includes a comprehensive tutorial to guide users through the setup process. ## Future Research Trajectory Wisteria represents an effort to move beyond the limitations of frontier models that lack native reasoning options. By enabling users to inspect reasoning text and manage memory locally, the project aims to provide a more stable and personal AI experience. The developer intends to continue stabilizing the architecture and refining research findings. ## Sources covering this story - https://negativeabsence.com/wisteria/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._